David Dalrymple, an AI safety specialist at the United Kingdom's publicly funded Aria research agency, warned that governments may run out of time to prepare for systems whose capabilities are improving faster than public safeguards. He pointed to a gap between technology companies and the officials responsible for energy networks and other infrastructure. The United Kingdom's AI Security Institute has reported rapid gains across several kinds of tasks. Dalrymple's program is therefore working on ways to contain failures in critical systems instead of assuming that advanced models will behave reliably.

Safety work follows a different clock from product development. A company can release a model, measure errors, issue an update, and keep its customers inside the experiment. A power grid cannot treat a blackout as useful feedback. Public agencies must test rare failures, assign responsibility, write procurement rules, and prepare operators before a system touches a control room. Those steps look slow beside a release calendar, yet their slowness records the number of people who cannot consent to the trial. Dalrymple's warning describes an institutional race in which one side earns money by moving first and the other is blamed after something breaks. The safeguard arrives as a budget request, a test environment, and an emergency procedure whose deadline was set by a company elsewhere.